Parental health shocks, child labor and educational outcomes: Evidence from Tanzania

Journal of Health Economics 44 (2015) 161–175

Contents lists available at ScienceDirect

Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase

Parental health shocks, child labor and educational outcomes: Evidence from Tanzania夽 Shamma Adeeb Alam ∗ Department of International Studies, Dickinson College, 28 North College St., Carlisle, PA 17013, United States

a r t i c l e

i n f o

Article history: Received 18 November 2014 Received in revised form 15 September 2015 Accepted 17 September 2015 Available online 28 September 2015 Keywords: Parental health Illness Shocks Education Child labor Tanzania

a b s t r a c t This paper examines the impact of parental illness on children’s education. We find that only father’s illness decreases children’s school attendance. Father’s illness also has long-term impacts on child education, as it decreases children’s likelihood of completing primary school and leads to fewer years of schooling. However, we find no evidence that father’s illness affects schooling through increased child labor. Instead, father’s illness decreases household’s income and reduces school attendance possibly because of the reduced ability of the family to afford education. In contrast, mother’s illness and illness of other household members have no effect on children’s schooling. © 2015 Published by Elsevier B.V.

1. Introduction Major illnesses are known to severely affect households; they can substantially increase medical expenditure and decrease household income and consumption for resource-constrained households in developing countries (Gertler and Gruber, 2002; Wagstaff, 2007; Genoni, 2012)1 . Consequently, households adopt a number of coping mechanisms, such as borrowing, dissaving, and

夽 This research is supported by grants from the William and Flora Hewlett Foundation (WFHF) Grant No.: 2012-7263 and the Institute of International Education (IIE). WFHF and IIE did not play any role in the design of this study. Any error in the paper is solely mine. I wish to thank the journal editor, Luigi Siciliani, three anonymous reviewers, Claus Portner, Rachel Heath, Hendrik Wolff, Mary Kay Gugerty, Brian Dillon, Willa Friedman, Subha Mani, seminar participants at University of Washington, Seattle, and conference session participants at 2014 Northeastern Universities Development Consortium (NEUDC) Conference, 2013 Pacific Conference for Development Economics, 2013 Population Association of America (PAA) Conference, 2013 Northwest Development Workshop, 2013 Western Economic Association International Annual Conference, 2013 Southern Economic Association Annual Conference for valuable comments and suggestions. ∗ Tel.: +1 7172548167. E-mail address: [email protected] 1 There is some disagreement in the literature about the effect of adult illness: Gertler and Gruber (2002) find a decrease in both household income and consumption following adult illness, but Genoni (2012) argues that only income is affected. On the other hand, Wagstaff (2007) finds that food consumption decreases and medical expenditure increases following adult illness. http://dx.doi.org/10.1016/j.jhealeco.2015.09.004 0167-6296/© 2015 Published by Elsevier B.V.

selling of assets, to mitigate the effects of such shocks (Wagstaff and Lindelow, 2014). Comparative studies suggest that major illnesses are as frequent, costly, and unpredictable as other income shocks like an unexpected crop loss, a decline in crop prices, or unemployment (Gertler and Gruber, 2002; Kochar, 1995; Wagstaff and Lindelow, 2014). Given the detrimental effects of illnesses, coupled with the long-term positive effects of schooling on adult earnings, it is important to understand how illnesses of adults, especially parents, influence child labor and schooling outcomes in a household. This paper explores the link between parental illness and child schooling. Prior literature provides evidence that economic shocks hinder schooling through the following two channels: (i) reducing household’s ability to afford education, and (ii) reallocation of children’s time from school to work (Beegle et al., 2006a; Dillon, 2008; Duryea et al., 2007; Jacoby and Skoufias, 1997). Parental illness may also affect child education through these two channels. (i) Income channel: Parents’ illness may decrease their own productivity at work or cause the parents to miss work entirely. For a credit-constrained household, the net income lost from missed work and increased medical expenditure may reduce the household’s ability to afford child education. (ii) Child labor channel: When a parent is ill, the household may need cheap labor to substitute for the parent’s missed work at the farm or in the household. Thus, even if the parents are able to afford child education following an illness, they may need their children to leave school to substitute for parents’ work in order to meet the household’s labor demands.

162

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

By employing longitudinal data with individual fixed effects from Tanzania, we examine whether parental illness affects child labor and schooling outcomes, and explore the channel through which the impact on schooling occurs. Using detailed time use data, we examine whether parental illness causes households to reallocate children’s time from school to work, and whether there are differential impacts from illnesses of fathers as compared to mothers. We find that only father’s illness hinders children’s education by decreasing their attendance. We also find strong suggestive evidence that father’s illness has long-term effects on children’s education, as it decreases their likelihood of completing primary school and ultimately causes them to finish fewer years of school. However, illness of fathers does not affect child labor. In contrast, illness of mothers has no effect on child schooling outcomes and only causes a small increase in child labor. Similarly, illness of other household members, such as grandparents, infants, child siblings, and adult siblings, has no effect on educational outcomes. Overall, we find no evidence that illness of parents and of other household members affects children’s schooling due to increased child labor. Instead, the results suggest that the effect on schooling may be primarily occurring through the income channel. As fathers are typically the primary income earners in households in Tanzania, only their illnesses substantially decrease household income and may consequently reduce the household’s ability to afford child education. These results are surprising in contrast to the literature on parental deaths. While some studies argue that both maternal and paternal deaths decrease schooling, others find that it is primarily maternal deaths that affect schooling2 . While the precise reason for the difference in the findings in the literature is unclear, the latter finding is consistent with another branch of literature, which shows that children work as substitutes for mothers. When mothers spend more time on market work, children, especially girls, spend less time in school and more hours working to substitute for mothers’ domestic duties (Hazarika and Sarangi, 2008; Katz, 1995; Skoufias, 1993). However, in our study, as we find that mother’s illness only causes a small increase in child labor (only 2 hours in prior week), it consequently does not displace time spent in school. We also find that parental illness does not have a differential effect by child gender on schooling. This is in contrast to findings in other countries that show that boys receive preferential treatment compared to girls, primarily due to intrahousehold resource allocation that favors boys (Deolalikar, 1993; Rose, 2000; Strauss and Thomas, 1995). Similarly, Pitt and Rosenzweig (1990) find that illness of infants is more likely to decrease schooling for a female sibling than a male sibling, as girls typically care for the infants in their families. However, we find no such biases in our estimations. Additionally, as boys and girls have a similar mean attendance rate in our study area, this suggests that there is no gender bias in primary schooling in this region3 . On the relationship between parental illness and child schooling, there has been only limited empirical evidence in the prior literature. Only one study, Bratti and Mendola (2014), establishes a causal relationship between parental illness and child schooling. They use panel data with individual fixed effects to find that

2 While Beegle et al. (2006b), Gertler et al. (2003) and Case et al. (2004) argue that the death of either parent hinders schooling, Ainsworth et al. (2005), Case and Ardington (2006) and Evans and Miguel (2007) find that maternal deaths have a much greater effect on schooling than paternal deaths. 3 Also, Thomas (1994) finds a gender bias by parents, noting that fathers typically invest more resources in their sons, while mothers invest more in their daughters. Other studies simply show that mothers typically invest more in children compared to fathers, especially with regard to children’s health (Case and Paxson, 2001; Case et al., 2000).

mother’s illness decreases secondary and tertiary school enrollment of older children (ages 15–24). In contrast, other studies – and our paper – focus on primary-to-middle-school aged children, ages 7–15. There are three reasons for this: (i) in many developing countries, most children are unable to continue schooling beyond primary school; (ii) a short-term negative impact on schooling at such young age can have long-term consequences on children’s schooling and labor market outcomes (Emerson and Souza, 2011; Beegle et al., 2006b); and (iii) labor work is considered to be child labor only until the age of 15. The few studies that do focus on primary or middle-school aged children find evidence that adult illness decreases schooling (Hannum et al., 2009; Sun and Yao, 2010) or increases the likelihood of child labor (Dillon, 2008; Bazen and Salmon, 2010). However, as these papers are based on cross-sectional data, and do not use instruments or panel data with fixed effects, they are unable to establish a causal relationship. Additionally, these studies have several data limitations: (i) first, they do not have time-use data, which is needed to understand the level of changes in children’s household work or transfer of hours from market work to household work and vice versa following parental illness; (ii) moreover, the lack of data on each individual household member’s illness (father, mother, child, or other household member) does not allow them to identify the effect of any specific individual’s illness on schooling (Hannum et al., 2009; Sun and Yao, 2010)4 ; (iii) and finally, the data from the Sun and Yao study may suffer from recall error, as individuals were asked to remember the timing of illness over the prior 15 years. Our paper, in contrast, uses a four-wave panel survey that includes a comprehensive time use survey for each household member and detailed information on health shocks. This paper contributes to the literature in several ways. This is the first paper that shows clearly, using individual fixed effects, that father’s illness can hinder child schooling; the only prior study with clear identification, Bratti and Mendola (2014), finds that only mother’s illness affects schooling. Moreover, to our knowledge, this is also the very first paper to use panel data with individual fixed effects to examine the impact of parental illness on schooling outcomes of primary and middle school aged children, an age-group for whom shocks can have long-term consequences on educational and labor market outcomes. Employing individual fixed effects, we are able to address concerns on unobserved heterogeneity that have biased cross-sectional estimates of prior studies. Additionally, while prior studies have focused only on parental illness, we are able to use detailed data on illness of each household member to specifically examine the impact of illness of siblings, grandparents, and other household members on children’s schooling outcomes. Furthermore, this work adds to the child labor literature (Edmonds, 2005). Ours is the first paper that uses detailed timeuse data to examine the effect of health shocks on child labor. Employing time-use data in a panel setting allows us to examine the reallocation of children’s time from school to work following parental illness, and thus examine a potential channel through which shocks may affect schooling5 . In contrast to prior studies, the time-use data allows us to examine separately the effect of illnesses on both household work and market work of children. Moreover, this paper contributes to the literature on intergenerational transfer of human capital. Prior studies have shown that parental socioeconomic status (SES) can affect a child’s health, which can consequently affect the child’s SES by hindering the child’s future educational and labor market outcomes (Currie, 2009;

4

Their survey does not identify which household member was ill. Several studies examine the tradeoff between child labor and schooling: Ravallion and Wodon (2000), Beegle et al. (2006a) and Janvry et al. (2006). 5

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

Case and Paxson, 2011; Strauss and Thomas, 1995). In this paper we explore a different channel: we examine whether parental health, which is likely to be partly determined by parental SES, affects the child’s educational outcomes, which is likely to subsequently affect the child’s SES. Finally, this paper adds to the literature on consumption smoothing and coping mechanisms adopted by households facing shocks (Kochar, 1999; Morduch, 1995; Townsend, 1994; Rosenzweig and Wolpin, 1993). Our results indicate that households decrease schooling to mitigate the effect of health shocks on consumption6 . It is important to note that while this paper and prior literature on shocks primarily focus on the income channel and child labor channel, parental illness may also affect schooling through other channels, such as psychological distress for the child; lack of parental supervision; shame related to HIV/AIDS of the parent; and a rising expectation that an ill child may die young, and therefore schooling investments not being worthwhile for the household. While we find evidence that father’s illness affects schooling through the income channel, it is possible that these other channels may also play a role in addition to the income channel. As the survey employed by this study does not specifically ask about these issues, we are unable to examine whether these additional channels affect schooling. The paper proceeds as follows: In Section 2 we discuss the data and describe the area where the survey data used for this study was compiled. The empirical methodology is presented in Section 3 and the results in Section 4. Finally, this study concludes with a discussion of the results and policy implications in Section 5. 2. Overview of data This study uses the Kagera Health and Development Survey (KHDS), a longitudinal survey conducted from 1991 to 1994 in the Kagera region in Tanzania. The Kagera region is in the Northwestern corner of Tanzania, west of Lake Victoria and bordering Burundi, Rwanda, and Uganda. It is a rural area primarily engaged in agriculture, with limited use of wage labor. In Kagera and other parts of rural Tanzania, there is a gendered division of labor within households. Typically, men are responsible for cash crop farming, the primary source of income for most households, as well as other income-generating activities, such as self-employment and wage labor. Women are usually in charge of household-related chores and the production of food crops in small garden plots (food crops are consumed by the household and typically do not generate income) (World Bank, 2009; Leavens and Anderson, 2011; CARE, 2010). While women’s primary focus is on food crops, they also spend time on cash crop production to varying degrees. Despite the significant labor contributions of women, men control nearly all household income (Rogers, 1983; Leavens and Anderson, 2011). The World Bank and the University of Dar es Salaam conducted the KHDS in four rounds from September 1991 to January 1994. The KHDS surveyed over 800 households in six districts of Kagera. The sampling was randomized based on the 1988 Tanzanian Census7 . Households were interviewed on a rolling basis throughout the year, with an average interval of six to seven months between each

6 This finding is similar to studies that show that income shocks, such as financial crises, natural hazards, crop losses, and unemployment, cause households to increase child labor and decrease school attendance to mitigate the effect of such shocks on consumption. The following papers find that income shocks decrease school attendance: Duryea et al. (2007), Fallon and Lucas (2002), Funkhouser (1999), Jacoby and Skoufias (1997), Jensen (2000) and Thomas et al. (2004). Furthermore, Duryea et al. (2007) and Beegle et al. (2006a) find that child labor increases following income shocks. 7 For further details on the sample selection, please refer to World Bank (2004).

163

survey round. The KHDS contains detailed information on all household members, such as illness suffered, education, hours worked, assets owned, and a number of other socioeconomic characteristics. Overall, 90.4% of all households remained in the survey for all four rounds. Probit estimations show that adult illness does not predict the incidence of households leaving the survey during the four survey rounds. Similarly, educational outcome of children also does not predict that an individual will leave the survey. When households dropped out of the survey, the surveyors randomly included additional households into the survey as replacements. In addition to the original four rounds, a fifth survey round was conducted in 2004. This fifth round is particularly useful for this study, as it allows us to follow up on the children 13 years after the first round and to examine the effect of parental illness on their final educational attainment. Similar to the prior four survey rounds, adult illness or child educational outcomes in prior rounds do not predict the incidence of households leaving the survey in this fifth survey round. For the remainder of the paper, only the male household head is referred to as the father, and the wife of the household head or a female household head is referred to as the mother. This is because a fraction of the children (8%) are orphans. Thus, the household heads are not the orphans’ biological parents but are actually their grandparents or uncles/aunts, and they play the role of parents to the orphans. To keep our terminology simple, we continue to refer to those parents as “father” or “mother.” 2.1. Education and child labor in Tanzania This paper focuses on children aged 7–15. This is because the minimum age for starting school is 7 (children are expected to start primary school when they are between 7 and 8 years of age), and it takes 7 years to complete primary school8 . The minimum age for legal employment is 15, and thus, children who are older than 15 and also work are not considered to be doing child labor. The survey contains detailed data on education and time-use of each household member aged 7 and above. Household members were asked the number of hours they worked, broken down by specific tasks, in the seven days prior to the interview. This allows us to determine the number of hours each individual spent weekly on household work and market work, along with the total hours worked9 . Additionally, KHDS compiles data on child’s education, such as, if they were ever enrolled in school, their school attendance in the last week, the highest grade they completed in each of all five survey rounds, and the time it took to reach school from home10 . The school year starts in January for primary schools in Tanzania. School expenses include the cost of attendance, textbooks, and uniforms, and these expenses increase with each higher grade level. The average annual schooling cost for a child in this sample was 1200 Tanzanian shillings (TZS). During the period of the study, the median household savings in Kagera was approximately 3000 TZS, and the median monthly per-capita income was approximately 2700 TZS. These income and savings statistics suggests that schooling represents a significant cost to households: 1200 TZS is

8 There is no separate middle-school in Tanzania. Children complete 7 years of primary school to move on to secondary school. In other words, middle-school aged children in the context of other countries are part of the primary school in Tanzania. 9 This study defines the following tasks as household work: preparing meals, cleaning the house, doing laundry, shopping for food, collecting firewood, collecting water, and caring for the ill. Market work is comprised of the following tasks: outside employment, work in the farm and garden (i.e., cash crop and food crop production), processing crops for sale, caring for poultry and livestock, collecting or transforming household livestock and animal products for sale (milk, eggs, etc.), attending to household business, and seeking additional paid work. 10 The median time to reach school from home is 20 min.

164

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

greater than or equal to the savings of 34% of all households. A study on Tanzania by Burke (1998) shows that 46% of parents report that they are unable to afford educational costs for their 13 to 15 year-old children. 2.2. Illness As we examine the effect of health shocks, it is important to note the major illnesses present in the region at the time of the survey. HIV/AIDS was prevalent in parts of Kagera. In the regional capital Bukoba in the Northeast, the adult HIV/AIDS rate was as high as 24% during the time of the survey (World Bank, 2004). However, in the southern and western regions of Kagera, the HIV/AIDS rates were nearly at zero. Malaria was also prevalent in this region: 32% of individuals with an illness in this survey reported that they believe they were suffering from malaria as they had severe malaria-type symptoms. However, only a fraction of these individuals visited any health facility. Additionally, 13% of individuals reporting an illness were diagnosed with malaria by a health practitioner. In overall Tanzania, malaria was the number one cause of visits to health facilities and the leading cause of deaths in hospitals in 1993 (Kavishe and Munshi, 1993). While there is no data on pneumonia and diarrhea/vomiting in the Kagera region, national data shows that these two diseases were respectively the second and third largest causes of deaths and visits to health facilities in Tanzania (Kavishe and Munshi, 1993). While KHDS does not have data on the types of illnesses suffered, it does contain detailed data on the incidence of illness during the 4 weeks prior to the respective survey rounds. KHDS asked all household members about any illnesses they had suffered in the 4 weeks prior to the interview, though these illnesses could have begun much earlier. Furthermore, the survey also asks “For how many days were you unable to conduct your usual activities because of this illness?” As there can be substantial variation in the severity of illness, this study only considers an individual to be ill if, a person reports illness and also reports they were unable to conduct their usual tasks for at least a day because of the illness11 . This measure of illness allows us to identify individuals whose physical functioning is affected, which then affects their efficiency or ability to work. This measure provides a similar objective measure as the Activities of Daily Living (ADL) used in other studies (Gertler and Gruber, 2002; Genoni, 2012), which ask individuals to rate if they are able to conduct a usual daily activity, such as the ability to stand up from a sitting position, to eat or bathe without help, or to walk uphill without assistance12 . In addition to questions on illness, all women aged 14–50 were also asked if they were pregnant at the time of the interview. An important question can be raised about the illnesses: are these illnesses short-term illnesses or chronic/long-term illnesses? While we do not have data on types of illnesses or diseases from the survey or from an external source, the data does suggest these are mainly short-term illnesses. The median length of illness among

11 There are variations in “daily activity” among individuals, for example across gender and age. Thus, individuals doing more physical labor may be more likely to be considered ill according to this definition of illness, because illnesses are more likely to have an effect on their daily activity. While individual fixed effects would address some concerns over these kinds of systematic measurement errors, this paper conducts robustness check by defining individuals as ill if a person only reports illness and the definition does not take into consideration whether the illness affected their daily activity. The estimates on the effect on schooling and child labor remain robust using this measure of illness. The results are presented in Table A5. 12 We also conduct robustness checks by using other definitions of illness that are based on limited daily activity. For example, household members were asked (i) whether the illness caused them to be confined to bed at any time; and (ii) whether the illness restricted them from doing their work. We find the results to be robust for both these definitions of illness. Results are presented in Tables A1 and A2.

Table 1 Summary statistics. Variable

Mean

Children who started primary schooling - Boys - Girls

70% 70% 69%

Attendance rate for children who started schooling - Boys - Girls

87% 87% 87%

Years of current schooling, if started schooling Final years of schooling in 2004

2.45

(2.0)

6.95

(2.0)

Proportion having finished primary school in 2004

84%

Percentage of children working

91%

Percentage of children with illness to their: - Any parent ill - Father ill - Mother ill - Adult sibling ill - Grand parents ill - Other household members ill - Children under 7 ill - Child’s own illness - Father facing change in health status over 4 rounds - Mother facing change in health status over 4 rounds

42% 21% 29% 14% 3% 7% 39.9% 39% 41% 55%

Mother pregnant

6.9%

Other household member pregnant

2.7%

Percentage of households having adult death

5.9%

Age

Std. Dev.

11.1

2.6

Number of total household members

7.5

(2.8)

Number of children

4.7

(2.1)

34%

(0.5)

Percentage of crop loss Household medical expenditure in each round (in TZS)

695

Per capita asset owned × 100 (in TZS)

1655.8

Number of observations

2590

(3624) (20,161)

This table provides the mean over all four rounds of survey unless otherwise noted. $1 = 526 Tanzanian Shillings (TZS) in 1994.

parents in the survey is 14 days and the mean length is 22 days. 18% of all parents reported suffering from chronic illnesses. However, only 11% of the parents who reported that the illness hindered them from their daily activity (this study’s definition of illness) also reported the illness to be chronic. So, there is only a small overlap between the illness defined in this study and chronic illness reported by individuals. However, it is important to note that these short spells of illness could be proxies for recurring illnesses suffered by individuals. 2.3. Summary statistics Table 1 presents the descriptive statistics. While 91% of all children reported working at least 1 h in the prior week, only 70% of children had started schooling. The children who had started schooling had an attendance rate of 87% in the prior week. There is no significant gender difference in the percentage of children who have started schooling (70% of boys and 69% of girls). Likewise, the attendance rate for boys and girls is equal in magnitude (87% for both boys and girls). Other studies in Tanzania also report similar findings. A nationally representative survey in Tanzania, conducted by the World Bank in 1993, also finds that the primary school enrollment ratio is similar for boys and girls (World Bank, 1999). The illness data shows that on average in each survey round, 42% of the children have at least one ill parent. 21% have an ill father, and

S.A. Alam / Journal of Health Economics 44 (2015) 161–175 Table 2 Summary statistics of hours worked.

Hours worked by: Children aged 7–15 - Boys - Girls Fathers Mother Other adult males Other adult females

Total hours

Household work

Market work

18.5 (13.1) 17.1 (12.2) 19.9 (13.7) 31.2 (23.9) 39.3 (22.2) 31.5 (22.9) 32.1 (23.0)

11.8 (9.2) 10.2 (8.5) 13.7 (9.5) 4.7 (8.2) 20.5 (13.6) 6.5 (8.4) 15.5 (13.1)

6.6 (7.9) 6.9 (8.1) 6.2 (7.6) 26.4 (21.8) 18.8 (15.1) 25.1 (21.1) 15.7 (15.3)

This table provides average number of hours worked in the prior week over four survey rounds. Standard errors are provided in parenthesis. Household work is defined as the following tasks: time spent preparing meals, cleaning the house, doing laundry, shopping for food, collecting firewood, collecting water, and caring for the ill. Market work is defined as the following tasks: outside employment, work in the farm and garden, processing crops for sale, taking care of the poultry/livestock, collecting or transforming household livestock/animal products for sale (milk, eggs, etc.), household business, and time spent seeking additional paid work.

29% have an ill mother13 . Overall, 41% of fathers and 55% of mothers had a change in health status over four survey rounds. Because households do not have access to health insurance, there is significant out-of-pocket health expenditure for sick individuals. The average per-capita household medical expenditure in each survey round is 695 TZS, which includes hospital fees and pharmaceutical expenses. Given the median monthly per-capita household income is 2700 TZS, these are significant medical expenditures for many households. Table 1 also lists the illness rates of other members of the household14 . Additionally, 7% of mothers were pregnant during the time of the interview. The survey provides detailed data on age, number of household members, crop loss shocks faced by the household in the prior six months, and the value of household asset holdings, which include land, business equipment, and durable goods. Lastly, over 5% of households faced adult deaths between survey rounds. Table 2 presents the descriptive statistics on hours worked, divided into three categories: (i) total hours worked; (ii) hours spent on household work; and (iii) hours spent on market work15 . On average, children work 18.5 h each week, with 11.8 h spent on household work and 6.6 h on market work. On average, girls work (19.9 h) more hours than boys (17.1 h), because girls spend more time doing household chores. However, there is a much greater gender difference in work among parents: fathers, on average, work 31 h and mothers work 39 h16 . Such large gender differences in hours worked have been found in several prior studies in Tanzania (Leavens and Anderson, 2011; CARE, 2010; World Bank, 2009). Mothers contribute almost equally to household (20.5 h) and market work (19 h). In contrast, fathers work substantially more hours on market work (26.4 h) compared to only 4.7 h on household work. Similarly, we find that other adult males in the household work mostly on market work, while other adult females split their hours almost equally between household work and market work.

13 A concern may be that fathers are unemployed, and therefore fathers report illness to justify their unemployment. However, this is unlikely to be a concern in this study as 95% of the fathers in each round report working in their own farms. Thus, as they are employed, they have no have no reason to justify unemployment by reporting illness. 14 Although 29% of all adults report suffering from illness, only 12% report visiting a health facility for that illness. The definition of a health facility includes health centers, hospitals and dispensaries, which are typically operated by NGOs, government and private organizations. These facilities are not uniformly accessible; while there are multiple health facilities in some communities, a few have no health facilities at all. For the latter case, people have to travel farther to access a health facility. 15 Variation of the important dependent and independent variables over all four survey rounds are provided in Table A3. 16 The average hours worked may appear to be low because there is a strong seasonality in farm work.

165

However, the gender difference in total hours worked by the parents is not present for other adults; adult males work a total of 31.5 h and adult females work 32.1 h each week. 3. Empirical methodology and specification This paper uses a linear model (OLS specification) with individual (child) fixed effects to find the effect of health shocks on child labor and schooling outcomes17 . The individual fixed effects allow us to control for time-invariant characteristics associated with the individual (child), such as: religion, gender, child birth order effects, parental characteristics and preferences, general health status of both the child and the parent, gender differences between children, and any other unobserved time-invariant child and parent heterogeneity. Thus, this estimation procedure will address concerns on unobserved heterogeneity that could have biased the results. The following specification is used for the estimations: Yi,j,t = ˇ0 +



ˇ1,k Illnesskj,t + ˇ2 Illnessij,t

k

+ ˇ3 Xi,j,t + ıi + t + εi,t

(1)

where subscripts i, j, and t denote individual, household, and survey rounds, respectively. Y represents the dependent variable for the estimation equation, which is school attendance and hours spent at work. Illnessk is a set of dummy variables indicating the illness of the kth person in the household, where k represents the following individuals: father, mother, grandparents, adult siblings, child siblings, and any other household member. ˇ1 ,k , the coefficient of Illnessk , is our coefficient of interest. Illnessi is a dichotomous variable representing the illness of the child. X represents a set of control variables, which include: the value of household assets and dummy variables for: crop loss, pregnancy status of mother and other women in household, adult death, and orphanhood status of the child.  t represents survey round fixed effects; as attendance or child labor hours may have seasonal variations, this specification also controls for the month of interview. ıi represents the individual fixed effects. Standard errors are clustered at the household level. Only those children who have already started primary schooling are included in the estimations. This is because many children in Tanzania start their primary schooling at a later age than 7 years. A detailed study by the World Bank in 1993 shows the prevalence of delayed entry into primary school in Tanzania. They find the average age of entry into primary school is about 9 years (World Bank, 1999). We find similar evidence in our data, as the primary school entry rate for 7-year-olds is only 20%. However, the primary school entry rate quickly increases with age, and 70% of all children in our data have started primary schooling. The main reason children start primary school late is because of overcrowding in Tanzanian schools (World Bank, 1999). Overcrowded schools force school administrators to first offer admission to the oldest children, who have not yet begun school, and if open spots remain in the entering class, children aged 7 are then admitted (World Bank, 1999). Additionally, some households may also intentionally delay sending a child to school because they may want the child to contribute to household work. If children who have not yet entered school are included in the school attendance estimation, it will not be possible to control for the unrelated above-mentioned factors that determine when a child enters school (as these can be timevarying effects, fixed effects will not account for those factors). That is why we only focus on the school attendance of children

17 We also conduct robustness check using fixed-effects logit model for school attendance, and find the estimates to be robust using this technique. Results for fixed effects logit model are presented in Table A4.

166

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

Table 3 Effect of parental illness on children’s school attendance. Dependent variable

School attendance (1)

(2)

(3)

(4)

(5)

(6)

Specification

RE

FE

FE

FE

FE

FE

Father ill

−0.027 (0.022)

−0.043** (0.021)

−0.044** (0.022)

−0.048** (0.022)

−0.052** (0.022)

Mother ill Ill child Grandparents ill Child aged 5 or below ill Children aged 6–18 ill Adult siblings ill Adult death Assets owned No. of household members Orphanhood status Region-time dummy Number of observations

0.022 (0.017) −0.044** (0.017) 0.060 (0.039) −0.025 (0.023) 0.029* (0.016) −0.009 (0.023) −0.151*** (0.045) 0.008 (0.005) 0.003 (0.003) −0.042 (0.034) No 2590

0.017 (0.016) −0.035* (0.020)

−0.131*** (0.042)

0.017 (0.016) −0.038* (0.020) −0.005 (0.031) −0.001 (0.021) 0.017 (0.016) −0.035 (0.024) −0.130*** (0.043)

No 2590

No 2590

0.016 (0.017) −0.040** (0.020) −0.006 (0.027) 0.003 (0.022) 0.022 (0.016) −0.034 (0.023) −0.128*** (0.043) 0.004 (0.011) −0.008 (0.006) −0.140** (0.067) No 2590

0.018 (0.017) −0.042* (0.021) 0.001 (0.029) 0.003 (0.023) 0.024 (0.016) −0.036 (0.024) −0.133*** (0.046) 0.002 (0.012) −0.009 (0.007) −0.159** (0.074) No 2465

−0.047** (0.022) 0.018 (0.018) −0.044** (0.021) −0.005 (0.027) −0.002 (0.024) 0.020 (0.016) −0.037 (0.024) −0.134*** (0.045) 0.002 (0.012) −0.010 (0.007) −0.165** (0.072) Yes 2465

Note: Linear Model. Mean attendance rate is 87%. Columns 2–6 represent individual (child) fixed effects estimations. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within household clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for month of interview and the round of survey. Estimations in columns 1, 4, 5, and 6 additionally controls for illness of other household members, crop loss, pregnancy of mother, pregnancy of other women in household. Additionally, estimation in column 1 also controls for sex, age fixed effects, district fixed effects, and years of parental education.

who have already started primary school18 . Using this criterion, unless otherwise noted, this paper focuses on 1144 children from 566 households. One potential concern with the estimations is that there may be estimation bias resulting from discrepancy in measurement periods: illness data is collected for the 4 weeks prior to the survey and school attendance and time-use data is collected for the prior 7 days. Therefore, if a parent recovers from an illness 7 days prior to the survey (before school attendance is measured), will our estimations capture the effect of parental illness on school attendance? This depends on how parental illness affects schooling. If parental illness affects schooling only through the income-channel (i.e., reduced income leading to reduced ability to afford education), then recovery of the parent before the last week can still affect school attendance in the last 7 days, because the parent may still be unable to afford child education. In such a case, there is unlikely to be a measurement error. And if there is measurement error (where parental illness did affect school attendance earlier, but not in the last 7 days), it will lead to an underestimate of the impact of parental illness on school attendance. However, if parental illness affects child schooling through the child labor channel, then the child will be replacing time spent in school with work until the parent recovers. If the parent recovers before the last 7 days, the child is likely to return to school for that period. This is likely to lead to a measurement error, because the reported data will show that a child did not miss school in the last 7 days following parental illness, even though in reality school attendance was previously affected. Therefore, in this case as well, the fixed effects regression will underestimate the effect of parental illness on school attendance. 4. Results and discussion

attendance is represented by an indicator variable that is equal to 1 if the child attends school for at least an hour in the prior week and 0 otherwise19 . Column (1) uses the ordinary least squares (OLS) with random effects (RE) to find the effect on school attendance. The RE estimate shows that while a child’s own illness significantly decreases attendance, parental illness does not have a significant effect20 . Surprisingly, it also shows that illness of siblings increase the likelihood of school attendance. However, individual or household level unobserved heterogeneity, which the RE estimation is unable to control for, may be confounding these effects of illnesses. Thus, we address such concerns using an individual fixed effects (FE) model (this estimation technique was also used by Bratti and Mendola (2014) to address similar concerns on child schooling estimation) in column (2). For fixed effects, we start with a parsimonious model where we only examine the impact on child school attendance from health shocks of primary interest: father’s illness, mother’s illness, child’s own illness, and any adult death in the family21 . We also control for the month of interview and survey round in this estimation. The FE result in column (2) shows that father’s illness leads to a 4.3 percentage point decrease (a 5% decrease) in school attendance, and the effect is statistically significant. Mother’s illness has a positive coefficient, but the magnitude is small and the effect is not statistically significant. In contrast, adult death and a child’s own illness significantly decrease attendance. To explore the effect of the illness of other household members, we add a set of controls for illness in column (3) for each of the following groups: grandparents, children younger than 5, children aged 5–18, adult siblings, and any other household members. These variables do not have a significant effect on school attendance and, moreover, the coefficient of parent’s illness remains robust to the inclusion of these variables. In column (4), we add a set of controls: value of assets owned per capita, orphanhood status of child, number of household members,

4.1. Effects on school attendance Table 3 demonstrates the impact of parental illness on school attendance of children who have already begun schooling. School

18 If children on average are starting school at the age of 9, they should be finishing primary school between the ages of 16 and 17. Therefore we conduct a robustness check to examine the effect of parental illness on school attendance for children aged 7–17. All the results remain robust for this age group. These results are available on request.

19

Children currently on vacation are not included in the estimation. In addition to the variables mentioned in Section 3, the OLS estimation also controls for child’s sex, age fixed effects, area (district) fixed effects and parental years of education. These variables are not included in the fixed effects estimations because the individual fixed effects control for the effects of these variables. 21 The reason for starting with such parsimonious model and not including control variables such as assets, crop loss, number of household members, and pregnancies, is because it is possible that parental illness can affect these control variables. Therefore, to be conservative and to avoid over-controlling bias, we start with this model. 20

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

167

Table 4 Effect of parental illness on child labor hours using individual fixed effects. Specification: Dependent variable

RE Total hours (1)

FE Total hours (2)

FE Total hours (3)

FE Household work (4)

FE Market work (5)

Father ill Mother ill Ill child Grandparents ill Child aged 5 or below ill Other children aged 6–18 ill Adult siblings ill Adult death Assets owned per capita No. of household members Orphanhood status

−0.238 (0.720) 1.170* (0.693) −2.396*** (0.552) 2.547 (1.590) 1.822*** (0.799) −0.097 (0.550) −0.734 (0.796) −0.444 (1.562) −0.287 (0.244) −0.778*** (0.112) 0.322 (0.995)

−0.481 (0.987) 2.020** (0.801) −1.512** (0.702)

−0.448 (0.980) 1.939** (0.824) −1.622** (0.699) 1.310 (1.988) 0.912 (1.036) −0.136 (0.702) −1.807** (0.873) 0.335 (1.798) −0.250 (0.381) −0.146 (0.306) 3.541 (3.299)

−0.731 (0.696) 1.273** (0.534) −1.440*** (0.492) 1.319 (1.618) 0.373 (0.750) 0.011 (0.529) −1.189* (0.676) −0.827 (1.054) −0.105 (0.294) −0.153 (0.204) 2.057 (1.932)

0.283 (0.587) 0.666 (0.545) −0.182 (0.406) −0.009 (1.347) 0.538 (0.580) −0.146 (0.409) −0.618 (0.539) 1.162 (1.198) −0.144 (0.232) 0.007 (0.183) 1.484 (2.898)

Dep variable: mean value Std. Dev

18.4 (15.1)

18.4 (15.1)

18.4 (15.1)

12.6 (11.1)

5.9 (7.8)

Number of observations

2465

2465

2465

2465

2465

0.626 (1.720)

Note: Linear Model. Columns 2–6 represent individual (child) fixed effects estimations. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within household clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for month of interview and the round of survey. Additionally, estimations in columns 1, 3, 4, and 5 control for illness of other household members, crop loss, pregnancy status of mother and pregnancy status of other women. Moreover, estimation in column 1 also controls for sex, age fixed effects, district fixed effects, and years of parental education.

crop loss, pregnancy status of mother and other women in the household22 . We find that the coefficients of illnesses remain qualitatively similar and father’s illness continues to have a significant effect on school attendance. In columns (5) and (6), we conduct a number of robustness checks. Although this sample only considers children who have started schooling, it is possible that some children may have dropped out of school before the start of the survey, and consequently those observations may bias the estimates. To eliminate the possibility of such bias, column (5) excludes children who did not attend school in any of the survey rounds. The coefficients remain similar and father’s illness decreases attendance by 5.2 percentage points (a decrease of 6%). While the individual fixed effects controls for time-invariant characteristics, there may be time-varying region-level shocks that will not be captured by the fixed effects or the survey round dummies. To capture these region-level shocks, survey round dummies interacted with each district are added to the specification in column (6). There are 6 districts followed over the 4 survey rounds, thus a set of 23 time dummies are added to the specification to capture time-varying unobserved heterogeneity at the district level. As shown in column (6), the coefficient of parent’s illness remains robust to the inclusion of district-time dummies. 4.2. Effects on child labor Next, we examine the effect on child labor to find evidence of any reallocation of time from school to work following illnesses. As only father’s illness affects schooling, we are particularly interested in examining if this effect arises via an increase in child labor. The results of child labor estimations are presented in Table 4. Similar to the school attendance estimations, we start by using a random effects model to find the impact of illnesses on child labor. Result in column (1) shows that while father’s illness has no impact on child labor, mother’s illness significantly increases child labor by 1.2 h in the prior week. Similarly, illness of children under the age of 5 significantly increases hours worked by 1.8 h. To account for any unobserved time-invariant characteristics, we employ FE model for estimations in columns (2) to (5). In column (2), we use a

22 A potential concern can be that maternal illness may be correlated with pregnancy status leading to multicollinearity issues. Robustness check shows that the results remain robust when pregnancy variables are dropped from the estimations. Results available on request.

parsimonious model where we only examine the effect of father’s illness, mother’s illness, illness of the child, and adult death. Additionally, we control for month of interview and survey round. In the FE model, we find that the estimated impact of mother’s illness almost doubles to 2.02 h (approximately 11% of the mean h of child work) each week; the coefficient of father’s illness is negative (−0.48 h) but statistically insignificant. In column (3) we introduce all the control variables and find that the coefficients of parental illness remain robust23 . Columns (4) and (5) disaggregate hours worked by type into household and market work. The disaggregation reveals that the increased labor hours following mother’s illness is mostly driven by an increase in household work. In contrast, fathers’ illness causes only a negligible and statistically insignificant reallocation of children’s labor from household to market work. 4.3. Identification of the channel of effect To summarize the results of the previous sections, while father’s illness hinders schooling, there is no evidence that it causes a reallocation of children’s time from school to work24 . Thus, we find no evidence that the effect on school attendance is due to increased child labor. Instead, father’s illness may affect attendance via another channel: a decrease in household income and

23 A possible concern can be that we are not finding a substantial impact on child labor as individuals may be unable to report time accurately, which can underestimate results. To explore this reasoning, we examine the proportion of healthy prime-aged adults reporting only few hours of work, for example reporting total weekly work below 24 h. 19% of non-sick parents aged below 55 (i.e. prime-age adults; it can be difficult for old parents to work long hours) report working below 24 h in a survey round. 19% may appear to be a significant number, but it is important to note that during non-peak agricultural seasons both parents are likely to spend substantially less time (compared to the average) on market work, which can decrease total hours worked to below 24 h, especially for fathers. It is worth noting that a study by Beegle et al. (2006a) employs this KHDS data to find that crop loss significantly increases child labor hours by 6 h each week. Given that Beegle et al. (2006a,b) has been able to find an impact on hours worked using this same data, we can expect that, if health shocks actually have an impact on labor hours, we should be able to pick up some of that effect in our estimations as the data should be reliable despite some noise. 24 While mother’s illness does not decrease child school attendance through increased child labor, it is important to note that mother’s illness may still affect the quality of a child’s learning. Mother’s illness increases hours of work of children, which may reduce the amount of time the child spends studying outside of school or the amount of time that the child is able to rest.

168

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

Table 5 Impact of illness of household members on income, consumption, and savings. Dependent Variable:

Per-capita household income (1)

Per-capita consumption expenditure (2)

Per-capita value of livestock owned (3)

Per-capita value of household savings (4)

Father ill Mother ill Ill child Grandparents ill Child aged 5 or below ill Adult siblings ill Assets owned per capita

−10,577 (11,515) −8668 (7031) 7798 (5922) 7590 (8087) 21,932 (23,978) −935.1 (4618) 711.7 (1844)

−1987* (1097) 113.7 (869.3) 673.3 (763.4) 502.8 (2171) −971.1 (1125) 441 (1421) 418.6 (725.6)

−271.4** (99.6) 8.32 (78.6) 177.3 (152.4) 214.2 (326.1) 2.219 (109.9) −1.69 (105.3) 87.10 (53.56)

−483.8 (665.4) 142.1 (398.1) 281.2 (697.6) 1869 (1954) 772.2 (611.2) 551.6 (403.5) −249.8 (306.0)

Dep variable: mean value Std. Dev

19,044 (164,502)

29,999 (22,002)

1055 (2624)

3051 (36,051)

Number of observations

1629

1629

1629

1629

Note: Linear Model with household fixed effects. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within household clusters. ** at 5%; * at 10%. All estimations control for illness of other household members, pregnancy of mother, pregnancy of other household members, orphanhood status, number of household members, month of interview, and the round of survey.

consequently, the ability to afford education. Therefore, we explore whether father’s illness has a differential impact on household income, consumption, and savings, compared to illness of other household members (Table 5). The survey asks each household their total consumption expenditure and income for the prior 6 months. Furthermore, in each survey round, households are asked the value of their savings/cash holdings and livestock owned. Therefore, using the illness data from the prior survey round, we find the impact of individual illnesses on income, consumption, and savings in the following 6 months. In other words, we find the effect of illnesses on household wealth 6 or 7 months after the illness has occurred (i.e. effect of illnesses in current survey round on wealth level in the following survey round). It is important to note that if the household recovers quickly following adult illness, then we would be unable to capture its effect on these variables; hence, these estimates would provide an underestimate of the actual effects. In column (1), we find that both father’s illness and mother’s illness decreases per-capita household income (10,577 TZS and 8668 TZS, respectively). Although these effects are not statistically significant, the difference in magnitudes of effects of father’s illness and mother’s illness is not large. This may arise from reporting errors in household income. It is important to note that the reported average per-capita household income (19,044 TZS) is substantially lower than the average per-capita household consumption (29,999 TZS). This may suggest that income is being underreported in the data25 . There is a large body of literature that shows that income is typically underreported in surveys in developing countries (Deaton, 1997; Thomas et al., 1991; Ravallion, 2003; Atkinson and Brandolini, 2001; Deininger and Squire, 1996; Szekely and Hilgert, 1999; Morris et al., 2000)26 ,27 .

25 The large standard deviation of income (164,502 TZS) compared to expenditure (22,002 TZS), where mean expenditure is actually greater than income, suggests that the numbers in the income data may be more spurious compared to expenditure data. 26 These studies provide several reasons for reporting errors in income: respondent may be unwilling to disclose income; underreporting because of income tax concerns, especially by the comparatively richer households; if only one person in household is surveyed (in this case, the household head), then that person may not know the income of other household members; if a farmer or trader sells their outputs at different points in time, they may not be able to keep track of their revenue from all sales; people may not be able to recall the exact timing of their income; individuals involved in agriculture and trade mostly do not keep track of inflow and outflow of funds and may not understand the question properly when they are asked about profits during a certain period of time (in this case the prior 6 months). This is not only a developing country issue; the same problem is prevalent in developed countries as well. 27 A concern may be that outliers for certain households’ expenditure is driving the high mean expenditure. However, examining the numbers suggest otherwise.

Therefore, studies suggest using household expenditure as a proxy for income, because expenditure data is less prone to the problems of underreporting as income data (Deaton, 1997; Deininger and Squire, 1996; Morris et al., 2000)28 . This is also evidenced in our data as average expenditure is 58% greater than average income. It may also be the case that given the seasonality in agricultural earnings, income is a poor measure of household wellbeing at the time when the interview was conducted. Instead, consumption expenditure can provide us a better idea of household wellbeing as households may be able to smooth consumption following shocks. Therefore, by examining the impact on household expenditure, we not only have a proxy for income, but this also informs us whether a household is able to smooth consumption following health shocks. Thus, in column (2) we examine the impact of parental illness on per-capita consumption expenditure, and find that father’s illness significantly decreases consumption (1987 TZS). Given that the average annual cost of schooling for a child is 1200 TZS, the average decrease in consumption expenditure is large enough to hinder an income-constrained household from affording a child’s education. We find that illness of other household members, including mothers, has no effect on consumption expenditure. To examine whether the decrease in household consumption affects a family’s liquid assets, we examine the effect on their livestock holdings and savings. We find that the total value of their livestock holdings decreases following father’s illness, which suggests that households may be selling their livestock to cope with this health shock (column 3). While father’s illness decreases savings, it does not have a statistically significant effect (column 4). Illness of other household members does not have a statistically significant impact on savings or value of livestock owned. These results may suggest that it is primarily father’s illness that decreases household income, and consequently, the household’s ability to afford child education. This set of estimations may also explain why illness of other household members, who may possibly be engaged in market work, has no effect on school attendance. For example, we find that illness of adult siblings has no impact on income. This may be because the average age of adult siblings is 26 years, and therefore illnesses of these young adult siblings may not be as severe as their parents, and

The maximum per-capita household income reported in the survey is 1233,598 TZS and the maximum per-capita household expenditure is 177,162 TZS. 28 It is important to note that the expenditure data may suffer from some of the same under-reporting issues as the income measurement, such as recall error and imputing biases. However the consensus is the level of under-reporting is much lower in expenditure data compared to income data.

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

169

Table 6 Effect of household member’s illness on hours worked by parents and other adults. Mother’s work

Father’s work

Dependent variable

Total hours (1)

Market work (2)

Father ill Mother ill Own illness Grandparents ill Child aged 5 or below ill Other children aged 6–18 ill Adult siblings ill

−0.027 (1.356)

−0.601 (0.941)

−6.577 (1.184) 0.015 (3.228) −0.122 (1.506) 1.053 (1.204) 1.767 (1.676)

Dep variable: mean value Std. Dev Number of observations

Other adult’s work

Total hours (3)

Market work (4)

Total hours (5)

Market work (6)

−3.299 (0.833) 0.972 (2.403) 0.090 (1.050) −0.052 (0.806) −0.360 (1.239)

0.248 (1.292) −5.919*** (1.281) −2.628 (3.978) −0.163 (1.459) 0.491 (1.264) 2.147 (1.813)

−0.330 (1.185) −5.657*** (1.168) 0.021 (3.969) 0.118 (1.358) 0.184 (1.187) 1.489 (1.718)

1.984 (1.733) 1.844 (1.422) −3.248* (1.811) 3.727 (2.908) −0.104 (1.821) 0.959 (1.341) −1.793 (1.661)

0.591 (1.163) 0.151 (1.109) −2.197 (1.451) 1.819 (2.492) −0.847 (1.482) 1.337 (1.058) −1.218 (1.300)

39.3 (22.2)

18.8 (15.1)

31.2 (23.9)

26.4 (21.8)

31.9 (23.0)

20.1 (18.6)

2051

2051

1603

1603

1957

1957

***

***

Note: Linear Model with individual fixed effects. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within household clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for illness of other household members, crop loss, per-capita assets owned, number of household members, orphanhood status, pregnancy status of mother and other women, month of interview, and the round of survey.

thus have no impact, on average, on household income or school attendance. Similarly, we find that illness of grandparents has no effect on income or school attendance. The reason may again be related to age and work capacity. The average age of a grandparent is 63 years. Moreover, a grandparent works on average 14 h each week on market work, which is 38% less than the mean hours of market work for parents. These numbers suggest that it may be difficult for grandparents to contribute to arduous farm activity, and consequently they may have a relatively smaller contribution to farm productivity and household income. This may be a reason why grandparent’s illness does not significantly affect household income and school attendance. Given the cultural setting in Tanzania, it is not surprising that only father’s illness affects a household’s income, as fathers are in charge of most income-generating activities for the household. However, mothers also spend some time on income-generating activities, such as cash crop production29 . Therefore, it is important to understand the labor allocation decision within a household following father’s illness to determine whether a mother works more to compensate for the fewer hours of work performed by father following his illness. Table 6 examines this issue. We find that mothers do not work more, neither in terms of time spent on market work nor in total hours, following father’s illness (columns 1 and 2). However, mother’s illness decreases her own work by 6.6 h in the prior week. Similarly, columns (3) and (4) show that fathers do not work more following mother’s illness; but fathers considerably decrease their own work by 6 h, as 5.7 h is reduced from market work (a 22% decrease from the mean level of market work). These results show that while fathers substantially decrease their time spent on income-generating activities following own illness, mothers do not work more hours to substitute for fathers’ market work. These results show a gendered division of labor within the household. It also possibly suggests that mothers are unable or unwilling to substitute for certain skills that fathers provide in market work (similarly, fathers are unable or unwilling to substitute for mothers’ work following mothers’ illness). It is also possible that other adults in the household work more to compensate for the hours that the parents are unable to work due to illness (66% of the other adults are adult siblings). Results in column 5 show that other adults work approximately 2 h more following parental illness, but the effect is not statistically

29 While women spend 18 h each week on market work, this includes a significant time spent on food crop production, which does not generate income.

Table 7a Break-down of school participation rate. Participation rate (%)

Percent of children (%)

20 40 60 80 100

2.1 1.5 3.7 11.1 81.6

significant30 . Combined, all these results suggest that other household members (children, mother, and other adults) may not be working significantly more on income-generating activities to compensate for the impact of father’s illness on household income. To further examine whether father’s illness primarily occurs through the income channel, we examine a new variable: regular school participation31,32 . The survey asks the days in which a child has attended school in the prior week. School is open during the 5 weekdays in a week. We create a variable named “regular participation”, where we divide the number of days a child actually attends school by the total number of days they should be attending school (5 days). Therefore, if a child attends school all 5 days in the prior week, then the child has a 100% participation rate. If a child misses 1 day of school, then the child has an 80% participation rate. Table 7a provides a breakdown of participation rate. It shows that 81.6% of children currently attending school have a 100% participation rate. Using this measure, we investigate whether illnesses of father or other household members affect the participation rate of children who are currently attending school33 .

30 It may seem surprising that there is not a significant increase in hours worked by other adults following parental illness. Two reasons may possibly explain this lack of statistical significance: (i) It may be the case that other adults are already working at their full capacity, especially in a peak agricultural season. Therefore, these adults are unable to significantly increase their hours worked following parental illness. (ii) Some households have several adults in addition to the parents. When a parent is ill, one of the other adults may work more, while other adults may not change their hours worked. Such differing variations in hours worked following parental illness can lead to a small average effect on hours worked and greater standard errors. 31 We thank a reviewer for suggesting this estimation. 32 While participation and attendance may generally refer to the same thing, here we deliberately use the word “participation” to distinguish it from the attendance estimations, thus avoiding confusion in understanding these two different results. 33 The estimations are conditional on child school attendance. If we include children who are not currently attending school (i.e. did not go to school even for 1 day) in this selected sample, it would not be a measurement of regular school participation rate. Instead, this will just be another alternate measure of attendance used in prior attendance estimation in Table 3. That is not the purpose of this estimation, and that is why children not attending school are excluded from this estimation.

170

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

Table 7b Effect of parental illness on children’s school participation rate. Dependant variable

Table 8 Effect of parental illness on child labor and attendance by gender.

School participation rate (1)

(2)

Father ill Mother ill Ill Child Adult death

−0.010 (0.013) 0.013 (0.011) −0.076*** (0.014) 0.016 (0.024)

−0.008 (0.013) 0.012 (0.011) −0.078*** (0.014) 0.012 (0.023)

Dep variable: mean value Std. Dev

94% (16%)

94% (16%)

All other controls included Number of observations

No 2465

Yes 2465

Note: Linear Model with individual (child) fixed effects estimations. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within household clusters. *** indicates significance at 1% level. All estimations control for month of interview and the round of survey. Estimations in columns 2 controls for illness of grandparents, adult siblings, child siblings, and other household members, crop loss, per-capita assets owned, number of household members, orphanhood status, pregnancy status of mother and other women.

Why examine the effect of father’s illness on regular school participation of children? If a father’s illness affects school attendance only through the income channel, as suggested by the prior estimations, then we should not expect father’s illness to affect regular participation of children who are in school. This is because this select sample of children in this estimation are already attending school (and hence have already paid the school fees), and therefore it cannot be the case that these children are unable to participate because of the household’s inability to afford education. Thus, if father’s illness does lead to irregular participation among children, it has to be through some other channel other than the income channel. The results of the estimation are provided in Table 7b. The results show that illness of household members, including father’s illness, has no effect on regular participation of children. This result may suggest that as father’s illness affects schooling through the income channel, it only affects child school attendance (estimations in Table 3) and does not affect participation rates in school. 4.4. Heterogeneity by gender and age of children We also explore if parental illness has differential effects by gender of parents and children (Table 8). The estimates show that parent’s illness does not have a differential effect (i.e., statistically insignificant) by child gender on attendance and child labor. This is not surprising in the context of Tanzania, as the mean level of child labor and attendance of boys and girls are not significantly different in magnitude. This is in contrast to the literature on gender differences in other countries, which finds that girls work

Dependent variable:

Attendance (1)

Hours worked (2)

Father ill × girls Mother ill × girls Father ill Mother ill Ill child Adult death Dependant variable: mean Number of observations

0.001 (0.043) −0.005 (0.031) −0.052* (0.027) 0.020 (0.022) −0.042* (0.021) −0.134*** (0.046) 87% 2465

0.459 (1.611) 0.463 (1.442) −0.646 (1.159) 1.737* (1.001) −1.626** (0.697) 0.358 (1.794) 18.4 2465

Note: Linear Model with individual (child) fixed effects. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within district clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for illness of grandparents, adult siblings, child siblings, and other household members, crop loss, per-capita assets owned, number of household members, orphanhood status, pregnancy status of mother and other women, month of interview, and the round of survey.

more and receive less schooling compared to boys. Consequently, when there is a greater demand for household labor, girls are preferred for work over boys, and girls’ labor hours come at the expense of their schooling (Pitt and Rosenzweig, 1990). In this paper, as there is no gender differential in attendance at the mean level, and given that father’s illness does not affect school attendance through increased child labor, it is not surprising that we find no differential effect by gender on children’s schooling. We also explore if the effect of parent’s illness varies with the age of children. As there is not enough statistical power in the interaction term of children’s age and parent’s illness, children’s age is divided into four groups: 7–9, 10–11, 12–13, and 14–15. Table 9 shows that mother’s illness does not have a differential impact on child labor or attendance for different age groups, and the results are not statistically significant. Similarly, the effect of father’s illness on school attendance does not vary with the age of the children, and the results are statistically insignificant. However, we find that when the father is ill, the oldest group of children, aged 14–15, is significantly likely to work increased hours (approximately 6 h more) compared to the youngest group. At the mean level, the oldest group of children is likely to work 2.4 h more following father’s illness. It is worth noting that the youngest group (aged 7–9) actually reduces their work by 3.4 h following father’s illness. This may suggest a labor complementarity/substitutability issue across age groups following father’s illness. Younger children may act as complements to father’s work, and may need parental supervision in tasks, which may explain why they work fewer hours following father’s illness. The oldest group of children is differentially affected, and may be acting

Table 9 Differential effects of parental illness on child labor and attendance by different age group. Dependent variable Father ill × age 10–11 ×Age 12–13 ×Age 14–15 Mother ill × Age 10–11 ×Age 12–13 ×Age 14–15 Father ill Mother ill Dep variable: mean value Std. Dev Number of observations

Attendance (1) −0.001 (0.069) 0.002 (0.070) −0.015 (0.072) −0.030 (0.059) −0.006 (0.058) −0.051 (0.061) −0.042 (0.061) 0.040 (0.053) 87% 2465

Hours worked (2)

Household work (3)

Market work (4)

2.689 (2.425) 2.018 (2.062) 5.839*** (2.171) −0.495 (2.161) −0.442 (2.025) −2.029 (2.155) −3.409** (1.720) 2.752 (1.751) 18.4 (15.1)

0.474 (1.567) −0.088 (1.441) 2.444 (1.518) 0.339 (1.455) 0.351 (1.382) −1.148 (1.520) −1.455 (1.138) 1.351 (1.166) 12.6 (11.1)

2.216 (1.373) 2.105* (1.212) 3.394** (1.364) −0.834 (1.266) −0.793 (1.166) −0.882 (1.236) −1.955* (1.027) 1.401 (1.006) 5.9 (7.8)

2465

2465

2465

Note: Linear Model with individual (child) fixed effects. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within district clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for illness of grandparents, adult siblings, child siblings, and other household members, crop loss, age-group dummy, per-capita assets owned, number of household members, orphanhood status, pregnancy status of mother and other women, month of interview, and the round of survey.

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

171

Table 10 Effect of parental illness on years of further education and likelihood of finishing primary school.

Specification

Years of education OLS

Likelihood of finishing primary school OLS

Years of education HH-FE

Likelihood of finishing primary school HH-FE

Years of education HH-FE

Likelihood of finishing primary school HH-FE

Illness variables Father’s illness

Dummy −0.311** (0.151)

Dummy −0.041 (0.027)

Dummy −1.506*** (0.558)

Dummy −0.198** (0.100)

No. of periods −0.211** (0.094)

Mother’s illness

−0.494*** (0.159)

−0.026 (0.029)

−0.789 (0.513)

−0.085 (0.115)

No. of periods −1.992*** (−0.515) −0.594 (0.386)

Dep variable: mean value Std. Dev

5.57 (2.5)

84%

5.57 (2.5)

84%

5.57 (2.5)

84%

Number of observations

843

843

843

843

843

843

Dependent variable

−0.043 (0.072)

Note: Linear Probability Model. Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within household clusters. *** indicates significance at 1% level; ** at 5%. All estimations control for illness of grandparents, adult siblings, child siblings, and other household members, age, gender, and orphanhood status.

as substitutes to father’s work, possibly because they need less supervision and have the most physical strength, allowing them to work more hours compared to younger children. We find a similar substitutability/complementarity issue in household work and market work (columns 3 and 4). For market work, all age groups other than the youngest work more hours following father’s illness (the result is statistically significant for age groups 12–13 and 14–15). The oldest group of children, ages 14–15, has the biggest increase in market work, an increase of 3.4 h, compared to the youngest group. For household work, we find again that the oldest group has the greatest increase in hours worked (2.4 h), but the effect is not statistically significant. The youngest group (ages 7–9) has a decrease in both household work and market work following father’s illness. Even though children aged 14–15 work more following father’s illness, their school attendance is not differentially affected. This may be because the mean increase in work for the oldest group is only 2.4 h following father’s illness; this may not be a substantial enough increase in hours worked to cause children to spend less time in school. This result suggests that father’s illness may not be affecting attendance via increased child labor. 4.5. Long-term effects on education It is important to note that school attendance only informs us about the short-term schooling status of children. Since children may reenroll after temporarily leaving school, it is important to understand the potential longer-term effects on education. Thus, this paper examines the impact of parental illness during the period 1991–1994 on educational outcomes achieved between 1994 and 2004. More specifically, we examine the impact of parental illness on: (i) the total years of schooling children have completed between 1994 and 2004 and (ii) whether they finish primary school by 2004. As primary education is 7 years long, children enrolled in school in 1994 are expected to finish their primary education by 2004. As we are limited to one observation for each individual, there is no within-individual variation. Thus we are unable to use individual fixed effects for these estimations. As we examine the impact of prior history of parental illness on child schooling levels in 2004, we have to modify our measurement of illness. To compare the effect of illness between 1991 and 1994 to the cases of no illness, we create new dummy variables that capture illnesses: the dummy variables now equal 1 if a family member suffered from an illness at least once between 1991 and 1994 and the child was of school age, and 0 otherwise. Only children who were between the ages 7–15 in any of the survey rounds are included in this estimation. Columns (1) and (2) in Table 10 present the OLS (cross-sectional) estimates, which show that both father’s and mother’s illness cause a significant decrease in total years of schooling. However, neither

parent’s illness has any effect on the likelihood of children finishing primary school. As discussed earlier, OLS estimations are unable to control unobserved heterogeneity that may affect both the likelihood of parental illness and children’s future education. To control for unobserved household time-invariant characteristics, household fixed effects are introduced into the specification. How do we get variation in the impact of parental illness on schooling with household fixed effects when all siblings in the same household are likely to have the same exposure to parental illness? The variation arises from the fact that a case of parental illness can only affect schooling decision when a child is eligible to be in school (ages 7–15); if a child is ineligible for school (below the age of 7), parental illness simply cannot have an impact on their schooling. Therefore, when children below the age of 7 become eligible for school in a later survey round (when they turn 7), we can have siblings in a household with two types of exposures: siblings whose schooling were affected by parental illness (as their age was between 7 and 15 when the parent was ill) and siblings whose schooling were not affected by illnesses (as they were not eligible for school during the parental illness) in the same sample. Thus, we can use the varied exposure of children’s schooling outcomes to parental illnesses for our estimations34 . Illustrating this estimation with an example: suppose we have two siblings, aged 8 years and 5 years respectively, who suffer from their father’s illness. For this estimation, the dummy variable for father’s illness is 1 only for the 8 year-old. As the 5 year-old is not eligible for school, the illness cannot affect schooling. That is why the father’s illness variable will be 0 when the 5 year old becomes eligible for school two years later (when he/she turns 7) and becomes part of the estimation sample of children aged 7–1535 . Therefore a child who is in our sample for more survey rounds may have more exposure to parental illness compared to his or her sibling who is in our sample for a shorter time (resulting from the age cutoff). Columns (3) and (4) present the results with household fixed effects. We find that father’s illness causes a 1.5 years decrease in years of schooling. This effect is statistically significant and substantially greater than the effect found in the OLS estimate (0.3 years). Father’s illness also significantly decreases the likelihood of finishing primary school by 20 percentage points (a decrease of 24%).

34 A question can be: why are we including children ineligible for primary school in this estimation when our focus has been on children aged 7–15? For this estimation, we only include children who were between the ages of 7–15 at least in one survey round. This allows us to include children previously ineligible for school (below age 7), but became eligible for school in a later survey round (i.e. reached the age of 7). 35 There may be a concern that there can be potential long-term indirect effects on the 5-year old, for example if the parental health shock is very severe, it may affect their schooling outcome two years later. However both siblings would suffer from that indirect effect, and the 8 year old’s schooling is likely to suffer more as he/she is eligible for school for the two intermediate years.

172

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

However, consistent with previous results, mother’s illness does not have a statistically significant effect on these long-term educational outcomes when the household fixed effects are employed36 . The current dummy-variable measurement of illness does not take into account potential variations in effects of illnesses that may be present for more than one survey round. To capture the effect of illnesses in multiple rounds, we create new illness variables: the number of survey rounds that individuals were sick between 1991 and 1994 when a child was of school age. Using these new illness variables, we examine the impact of number of periods of illness on schooling completed till 2004. Results in columns (5) and (6) show that father’s illness in an additional survey round leads to a decrease of 2 years of education and decreases likelihood of finishing primary school by 21 percentage points37 . These coefficients are statistically significant. Similar to prior results, mother’s illness does not have a statistical significant effect on these schooling outcomes. Although the above specifications do not allow the use of individual fixed effects, they control for several important individual characteristics, such as gender, age, orphanhood status, and the number of periods the children were in the sample. An important limitation of this analysis is that this estimation is unable to control for future shocks that occurred between 1994 and 2004 survey rounds. An example of a future shock can be a recurring illness for the parent. Such future shocks may be correlated with past parental illness, and hence bias the estimates if they are not controlled for in the estimation. Another caveat to this analysis of long-term effects is that as the analysis is dependent on the timing of entry/exit of siblings from the sample, the identification of the effect of illnesses is dependent on the birth spacing of the children. Therefore, Table 10 only provides suggestive evidence for the effect of parental illness on final education level. It is also important to note that the retention rate of respondents from the initial 1994 survey round to the final 2004 survey round is 82%. Hence, attrition may bias these estimates on long-term educational outcomes. 4.6. Effect on the likelihood of starting primary school It is possible that parental illness may cause a delay in a child’s first-time entry into primary school (i.e. start of schooling). We examine whether there is any such effect in Table 1138 . For this estimation, it is important to note three timing issues that make the estimation challenging: (i) households are, at most, interviewed twice in a given year in 6/7 months intervals and that is when we learn from the data whether a child started schooling; (ii) households are only asked about illnesses in prior 4 weeks, and therefore we do not have individual illness history beyond the 4 weeks; and (iii) children can enter a school only at the beginning of an academic year, in January. Therefore, an entry to school in January may only be affected by illnesses few months prior to January. As we do not have illness data beyond prior 4 weeks, we cannot examine the impact of illness on school entry for households that were interviewed any time between March–December. This is because if a household was interviewed in March, we do not have their illness record prior to the school entry decision in January of the same year. Only for the interviews in the months of January and February, we may have illness data from prior 4 weeks that were before the school entry decision.

36 We conduct robustness check for other age groups, ages 7–17 and ages 9–17, as children typically start school late in Tanzania. The results remain robust for these age groups and are available on request. 37 As 73% of fathers with illnesses have the illness only for a single survey round, the magnitude of the coefficient of father’s illness is not substantially different than when a dummy variable is used to measure illness in columns (3) and (4). 38 We thank two reviewers for suggesting to conduct this estimation.

Table 11 cross-sectional estimates of the effect of parental illness on likelihood of school entry. Dependent variable

Entry into primary school (1)

Father ill Mother ill Ill Child Grandparents ill Child aged 5 or below ill Children aged 6–18 ill Age Gender Dummy: female = 1 Father: Years of schooling Mother: years of schooling Dependent variable: mean Number of observations

(2)

−0.075 (0.041) −0.011 (0.043) 0.036 (0.050) 0.081 (0.116) 0.038 (0.051) −0.011 (0.040) 0.023*** (0.005) −0.005 (0.034) 0.010* (0.006) 0.011* (0.006) 33% 866 *

−0.103* (0.052) −0.034 (0.047) 0.065 (0.067) −0.021 (0.121) 0.055 (0.066) 0.023 (0.051) 0.041*** (0.009) 0.004 (0.041) 0.019** (0.008) 0.012* (0.007) 35% 567

Note: OLS (Random effects). Standard errors are in parentheses and are computed after correcting for correlation and heteroskedasticity within district clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for illness of adult siblings, and other household members, crop loss, per-capita assets owned, number of household members, orphanhood status, pregnancy status of mother and other women in household, sex, district fixed effects, month of interview, and the round of survey.

Data from only two months (January and February) in a year clearly would not provide us enough observations to conduct an estimation. That is why, to utilize more illness data prior to school entry, we examine the effect of illness in prior survey round (t − 1) on school entry in current round (t). Illustrating the estimation with an example: for a household that was interviewed in March and asked about a child’s school entry decision for the year, their prior survey round would be in September/October. An illness in September/October may affect school entry decision in January, which would then be reflected in household responses to the interview in March. Similarly, for a household interviewed in April, if we use their illness data from prior survey round 6/7 months ago in October/November, it is possible that an illness in October/November can have an impact on school entry decision in January. However, we are unable to use the school entry observations for the months of September–December, because their corresponding lagged illness data were collected (in the months February–June) after the school entry outcome in January. Therefore, we do not have the lagged illness data before school entry for those months. For example, for a household interviewed in September, their lagged illness data would be in February/March, which is after the school entry outcome. Ideally, we would want to use a fixed effects specification for this estimation. However, we lose observations in four ways for this estimation: (i) focusing only on children who did not start schooling limits the data; (ii) as we use lagged illness data, we are unable to use data from first survey round; (iii) we are unable to use schooling data for September–December; and (iv) once a child starts schooling, their following school entry observations are not included in the estimation (this is because we are only trying to examine the likelihood of joining school). Dropping these observations mean that we do not have enough children with two or more observations to conduct an individual fixed effects estimation. That is why we use OLS random effects (RE) estimation strategy for this estimation. This leads us to our empirical strategy for this estimation: Yi,j,t,m = ˇ0 +



ˇ1,k Illnessk j,t−1 + ˇ2 Illnessi j,t−1

k

+ ˇ3 Xi,j,t + Ai + t + εi,t

(2)

where m {January, February, March, April, May, June, July, August}

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

173

Table 12 Examining the endogeneity of illness. Dependent variable: Assets before parental illness Crop loss prior to parental illness Dep variable: mean value Number of observations

Parental illness 0.003 (0.012) 42% 3520

Parental illness

Father’s Illness

Mother’s Illness

0.002 (0.012) −0.018 (0.026) 42% 3520

0.013 (0.017) −0.049 (0.035) 21% 1596

−0.009 (0.017) 0.010 (0.035) 29% 1924

Note: Linear Probability Model with individual fixed effects. Standard errors are in parentheses. Standard errors are computed after correcting for correlation and heteroskedasticity within household clusters. All estimations control for number of household members, month of interview, and the round of survey.

where m represents the month of interview when we receive the school entry information, and it is chosen such that the illness data in prior survey round was collected before school entry decision. Ai represents child level time-invariant controls employed, such as sex of the child and area (district) fixed effects. t represents current survey round and t − 1 represents prior survey round. Y represents a dummy variable, which equals 1 if a child joins the primary school, and 0 otherwise. All other variable definitions remain the same as before. The RE results are presented in column (1) of Table 11. The estimations suggest that father’s illness significantly decreases the likelihood of a child entering a primary school by 7.5 percentage points. Mother’s illness does not have a significant effect on school entry. Interestingly, both parents’ years of schooling seem to play an important role as greater parental education leads to a significantly higher likelihood of a child entering a primary school. One concern may be that for households that are interviewed for their school entry data in February, their lagged illness data (6/7 prior months prior) came from July/August. Illnesses, such a long time prior to January, may not affect the school entry outcome. Therefore as a robustness check, we only focus on illnesses few months prior to January school entry outcome. Thus, we include only lagged illness data for the four months prior to school entry: October to January (instead of July–January). These results are presented in column (2). It shows that the magnitude of the effect of father’s illness increases to 10.3 percentage points, and the effect is statistically significant. Illness of mother continues to have no impact on schooling. It is important to note that these RE estimations in Table 11 only provide suggestive evidence as they do not control for certain individual and household specific characteristics. Additionally, this estimation also does not control for spatial factors such as capacity constraints in local school.

4.7. Examining the endogeneity of Illness In Table 12, we investigate potential endogeneity concerns: whether prior assets owned or income shocks cause parental illness. If health shocks are exogenous, we would expect that assets owned or income shocks to be orthogonal to illness. The following estimations test if assets owned and income shocks, measured in the form of crop loss, are determinants of parental illness. Column (1) shows that assets owned prior to the health shocks have no significant effect on parental illness. Similarly, crop loss before the prior survey round has no effect on illness (column 2). Lastly, we examine the effect specifically on father’s illness (column 3) and mother’s illness (column 4) and find that neither prior crop loss nor assets owned are significant determinants of illness. This suggests that illness is not endogenously determined by a household’s wealth status or income shocks. We also conduct other sensitivity analysis and find that coefficients of parental illness remain robust to the exclusion of important control variables,

such as assets owned, crop loss, and illness of other household members39 . 5. Discussion and conclusion Employing longitudinal data from the Kagera region in Tanzania, this paper examines the role of illness of parents and of other household members on child labor and schooling outcomes. While prior study has shown that mother’s illness affects child schooling (Bratti and Mendola, 2014), this is the first study that clearly identifies, using individual fixed effects, that father’s illness also hinders child schooling. This is also the very first paper that uses individual fixed effects to examine the impact of parental illness on schooling for primary and middle school aged children (ages 7–15), a group for whom shocks can have a long-term consequence on educational outcomes. This paper also adds to the literature by addressing several questions not examined in prior literature, such as: does illness of siblings or other household members affect child schooling? Can the impact of parental illness on schooling occur through increased child labor? Or does the effect occur through a decrease in household income? Is there a gender or age-specific effect following parental illness? Can parental illnesses have a long-term impact on child schooling by affecting the total years of education attained? We find that although father’s illness hinders child education, there is no evidence that it causes a reallocation of children’s time from school to work. Therefore, father’s illness does not affect attendance through increased child labor. Instead, the results suggest that since fathers are typically the primary income earners in households, their illness decreases household earnings and may be decreasing child schooling because of their reduced ability to afford education. The paper also finds suggestive evidence that father’s illness affects children’s long-term schooling outcomes by reducing the children’s likelihood of finishing primary school and causes them to finish fewer years of school overall. In contrast, mother’s illness and illness of other household members has no effect on schooling. Surprisingly, we find no age or gender specific effect following parental illness. While the finding that father’s illness hinders child schooling may seem surprising, it is in line with some of the studies in prior literature. For example, Duryea et al. (2007) find that families are unable to afford child schooling following the unemployment of household heads in Brazil. Similarly, other studies have shown that income shocks decrease child schooling (Jacoby and Skoufias, 1997; Jensen, 2000). However, the findings of this paper are different from a number of prior studies on parental deaths (Beegle et al.,

39 We further analyze the effect of heterogeneity by female-headed households (i.e., single mothers). However, we find no significant effect of illness of single mothers. Illness of female household head also has no effect on child education because female heads do not work substantially more in market work than mothers in maleheaded households. Thus, their illness does not affect market income. These results are available on request.

174

S.A. Alam / Journal of Health Economics 44 (2015) 161–175

2006b; Gertler et al., 2003; Case et al., 2004; Case and Ardington, 2006; Evans and Miguel, 2007). Those studies show that maternal death, in addition to paternal death, hinders child schooling. The studies suggest that in addition to being a productive member of the household, mothers have the traditional role of child-rearing, which includes allocation of household resources to improve children’s human capital. Therefore, the absence (after death) of a mother leads to a decrease in allocation of resources towards a child’s health and education. A mother’s absence also leads to a loss of non-pecuniary parental inputs, such as role model effect, monitoring, and other important time spent by the mother in improving human capital or child quality. So, why does mother’s illness, unlike mother’s death, have no effect on child schooling in this study? This may be because when a mother is absent (after death), other members of the household are unable to substitute the permanent role of mother in providing the necessary human capital investments, non-pecuniary inputs, and child supervision. However, when a mother is ill, the mother is still present in the household to provide the non-pecuniary inputs and ensure that the child receives the necessary investment in their health and education. And in most severe cases of illness, the mother may at least be able to temporarily assign the role of child supervision to an older sibling. Given the findings of this study, policies are needed that insure children’s schooling against the risk of parental illness. Governments may consider providing financial assistance supporting schooling fees to schoolchildren who can provide medical evidence from health care providers that confirms the existence of parental illness40 . Vouchers providing financial support for books and uniforms could also be considered for children with parental illness. Schools may also consider allowing affected households substantial delays in the payment of schooling fees and expenses until after the household member recovers from illness.

Appendix A. Tables A1–A5 .

Table A1 Robustness check with other definitions of illness—effect of illness that confines household members to bed on children’s school attendance and child labor. Dependent variable

Father ill Mother ill Ill child Grandparents ill Child aged 5 or below ill Other children aged 6–18 ill Adult siblings ill Adult death Assets owned Number of observations

School Attendance

Hours Worked

(1)

(2)

−0.043 (0.023) 0.014 (0.019) −0.051* (0.027) −0.002 (0.038) 0.028 (0.024) −0.010 (0.019) −0.022 (0.028) −0.135*** (0.045) 0.004 (0.012) 2465 *

−0.160 (0.980) 0.919 (0.801) −2.420*** (0.874) 0.772 (1.951) 0.960 (1.002) −1.221 (0.819) −2.214** (0.949) 0.331 (1.761) −0.206 (0.378) 2465

Note: Linear model with individual (child) fixed effects. Standard errors are in parentheses. Standard errors are computed after correcting for correlation and heteroskedasticity within district clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for illness of other household members, number of household members, crop loss, orphanhood status, pregnancy status of mother and other women, month of interview, and the round of survey.

40

A concern is that not all households have easy access to health facilities.

Table A2 Robustness check with other definitions of illness—effect of illness that restricts household member from doing work on children’s school attendance and child labor. Dependent variable

Father ill Mother ill Ill child Grandparents ill Child aged 5 or below ill Other children aged 6–18 ill Adult siblings ill Adult death Assets owned Number of observations

School attendance

Hours worked

(1)

(2)

−0.052 (0.026) −0.020 (0.024) −0.109*** (0.034) 0.017 (0.045) 0.003 (0.022) −0.010 (0.018) −0.022 (0.028) −0.143*** (0.048) 0.003 (0.011) 2465 **

1.734 (1.262) 1.122 (0.856) −3.874*** (1.376) −0.944 (2.780) 0.817 (1.009) 0.529 (0.923) 0.167 (1.127) 0.698 (1.784) −0.154 (0.369) 2465

Note: Linear model with individual (child) fixed effects. Standard errors are in parentheses. Standard errors are computed after correcting for correlation and heteroskedasticity within district clusters. *** indicates significance at 1% level; ** at 5%. All estimations control for illness of other household members, number of household members, crop loss, orphanhood status, pregnancy status of mother and other women, month of interview, and the round of survey.

Table A3 Variation of important variables over four survey rounds. Overall average

Round of survey 1

2

3

4

Father’s illness Mother’s illness

21% 29%

18% 22%

23% 32%

23% 35%

19% 27%

Attendance rate - Boys - Girls

87% 87% 87%

85% 83% 86%

89% 91% 86%

87% 87% 87%

89% 89% 88%

Percentage of children working

91%

87%

90%

93%

93%

Hours worked - Boys - Girls

18.4 17.4 19.6

18.3 17.2 19.7

17.9 16.9 19

18.8 17.3 20.2

18.6 18.1 19.3

Household hours - Boys - Girls

13.1 11.7 13.5

12.7 11.9 13.6

12.3 11.4 13.2

12.4 11.4 13.3

12.9 12.2 13.8

5.4 5.7 6.1

5.6 5.2 6.1

5.7 5.5 5.9

6.5 6.0 6.9

5.7 6.0 5.5

Market work hours - Boys - Girls

Note: Above numbers are mean percentages.

Table A4 effect of parental illness on likelihood of school attendance using fixed effects logit model.

Father ill Mother ill Ill Child Child aged 5 or below ill

Coefficient (1) −0.841** (0.357)

Odds-ratio (2) 0.43** (0.15)

0.529 (0.323) −0.646** (0.314)

1.70 (0.55) 0.52** (0.16)

Other children aged 6–18 ill Adult siblings ill Adult death in past 1 year Assets owned Number of observations

−2.082*** (0.652) −0.070 (0.157) 404

0.125*** (0.08) 0.932 (0.15) 404

Coefficient (3) −0.989*** (0.384) 0.548 (0.340) −0.734** (0.325)

Odds-ratio (4) 0.372*** (0.14) 1.730 (0.59) 0.48** (0.16)

0.339 (0.456)

1.403 (0.64)

0.778** (0.357)

2.178 (0.78)

−0.786 (0.580) −2.296*** (0.696) −0.171 (0.176) 404

0.456 (0.26) 0.101*** (0.07) 0.842 (0.15) 404

Note: Standard errors are in parentheses. Standard errors are computed after correcting for correlation and heteroskedasticity within district clusters. *** indicates significance at 1% level; ** at 5%. Estimations control for illness of other household members, number of household members, crop loss, orphanhood status, pregnancy status of mother and other women, month of interview, and the round of survey.

S.A. Alam / Journal of Health Economics 44 (2015) 161–175 Table A5 Robustness check with other definitions of illness—defining individuals as ill if a person only reports illness but does not affect their ‘daily activity’–Effect of only illness on children’s school attendance and child labor. Dependent variable

School attendance (1)

Hours worked (2)

Father ill Mother ill Ill child Grandparents ill Child aged 5 or below ill Other children aged 6–18 ill Adult siblings ill Adult death Assets owned Number of observations

−0.045** (0.022) 0.028 (0.022) −0.032* (0.02) −0.004 (0.031) 0.027 (0.025) 0.021 (0.023) −0.030 (0.025) −0.137*** (0.046) 0.003 (0.014) 2465

0.589 (0.755) 1.243 (0.845) −0.71 (0.527) 0.136 (1.422) −0.366 (1.470) 0.895 (0.945) −1.513 (1.110) 0.621 (1.740) −0.259 (0.328) 2465

Note: Linear model with individual (child) fixed effects. Standard errors are in parentheses. Standard errors are computed after correcting for correlation and heteroskedasticity within district clusters. *** indicates significance at 1% level; ** at 5%; * at 10%. All estimations control for illness of other household members, number of household members, crop loss, orphanhood status, pregnancy status of mother and other women, month of interview, and the round of survey.

References Ainsworth, M., Beegle, K., Koda, G., 2005. The impact of adult mortality and parental deaths on primary schooling in North-Western Tanzania. J. Dev. Stud. 41 (3), 412–439. Atkinson, A., Brandolini, A., 2001. Promise and pitfalls in the use of “secondary” data-sets: Income inequality in OECD countries as a case study. J. Econ. Lit. 39 (3), 771–799. Bazen, S., Salmon, C., 2010. The impact of parental health on child labor: the case of Bangladesh. Econ. Bull. 30 (4), 2549–2557. Beegle, K., Dehajia, R., Gatti, R., 2006a. Child labor and agricultural shocks. J. Dev. Econ. 81, 80–96. Beegle, K., Weerdt, J., Dercon, S., 2006b. Orphanhood and the long-run impact on children. Am. J. Agric. Econ. 88 (5), 1266–1272. Bratti, M., Mendola, M., 2014. Parental health and child schooling. J. Health Econ. (34), 94–108. Burke, K., 1998. Investing in children’s human capital in tanzania: does household illness play a role. In: SUNY Stony Brook: Dissertation. SUNY, Stony Brook. NY. CARE Tanzania, 2010. Gender Analysis Report for CARE/Tanzania. CARE Tanzania. Case, A., Ardington, C., 2006. The impact of parental death on school outcomes: longitudinal evidence from South Africa. Demography 43 (3), 401–420. Case, A., Paxson, C., 2001. Mothers and others: who invests in children’s health. J. Health Econ. 20, 301–328. Case, A., Paxson, C., 2011. The long reach of childhood health and circum-stance: evidence from the Whitehall II study. Econ. J. 121, 183–204. Case, A., Lin, I., McLanahan, S., 2000. How hungry is the selfish gene? Econ. J. 110 (466), 781–804. Case, A., Paxson, C., Ableidinger, J., 2004. Orphans in Africa: parental death, poverty, and school enrollment. Demography 41 (3), 483–508. Currie, J., 2009. Healthy, wealthy, and wise? Socioeconomic status, poor health inchildhood, and human capital development. J. Econ. Lit. 47, 87–122. Deaton, A., 1997. The Analysis of Household Surveys: A Microeconometric Approach to Development Policy. World Bank Publications, Baltimore, MD. Deininger, K., Squire, L., 1996. A new data set measuring income inequality. World Bank Econ. Rev. 10 (3), 565–591. Deolalikar, A., 1993. Gender differences in the returns to schooling and in school enrollment rates in Indonesia. J. Human Resour. 28 (4), 899–932. Dillon, A., 2008. Child labor and schooling responses to production and health shocks in Northern Mali. In: IFPRI Working Paper. Duryea, S., Lam, D., Levison, D., 2007. Effects of economic shocks on children’s employment and schooling in Brazil. J. Dev. Econ. 84 (1), 188–214. Edmonds, E., 2005. Child labor. In: Schultz, T.P., Strauss, J. (Eds.), Handbook of Development Economics, vol. 4. Elsevier, pp. 3607–3709. Emerson, P., Souza, A., 2011. Is child labor harmful? The impact of working earlier in life on adult earnings. Econ. Dev. Cult. Change 59 (2), 345–385. Evans, D., Miguel, E., 2007. Orphans and schooling in Africa: a longitudinal analysis. Demography 44 (1), 35–57. Fallon, P., Lucas, R., 2002. The impact of financial crises on labor markets, household incomes, and poverty: a review of evidence. World Bank Res. Observer 17 (1), 21–45.

175

Funkhouser, E., 1999. Cyclical economic conditions and school attendance in Costa Rica. Econ. Educ. Rev. 14, 31–50. Genoni, M.E., 2012. Health shocks and consumption smoothing: evidence from Indonesia. Econ. Dev. Cult. Change 60 (3), 475–506. Gertler, P., Gruber, J., 2002. Insuring consumption against illness. Am. Econ. Rev. 92 (1), 51–70. Gertler, P., Martinez, S., Levine, D., Bertozzi, S., 2003. Losing the presence and presents of parents: how parental deaths affect children. In: Working Paper. Hannum, E., Sargent, T., Yu, S., 2009. Poverty, parental ill health and children’s access to schooling in rural Gansu, China. In: Gansu Survey of Children and Families Papers. Hazarika, G., Sarangi, S., 2008. Household access to microcredit and child work in rural Malawi. World Dev. 36 (5), 843–859. Jacoby, H., Skoufias, E., 1997. Risk, financial markets, and human capital in a developing country. Rev. Econ. Stud. 64, 311–335. Janvry, A.d., Finan, F., Sadoulet, E., Vakis, R., 2006. Can conditional cash transfer programs serve as safety nets in keeping children at school and from working when exposed to shocks? J. Dev. Econ. 79, 349–373. Jensen, R., 2000. Agricultural volatility and investments in children. Am. Econ. Rev., Pap. Proc. 90 (2), 399–404. Katz, E., 1995. Gender and trade within the household: observations from rural Guatemala. World Dev. 23 (2), 327–342. Kavishe, F., Munshi, S., 1993. Nutrition-Relevant Actions in Tanzania. United Nations Administrative Committee on Coordination, Adelaide. Kochar, A., 1995. Explaining household vulnerability to idiosyncratic income shocks. Am. Econ. Rev., Pap. Proc. 85 (2), 159–164. Kochar, A., 1999. Smoothing consumption by smoothing income: hours-of-work responses to idiosyncratic agricultural shocks in rural India. Rev. Econ. Stat. 81 (1), 50–61. Leavens, M., Anderson, C., 2011. Gender and agriculture in Tanzania. In: EPAR Biref No. 134. Morduch, J., 1995. Income smoothing and consumption smoothing. J. Econ. Perspect. IX, 103–114. Morris, S., Carletto, C., Hoddinott, J., Christiaensen, L., 2000. Validity of rapid estimates of household wealth and income for health surveys in rural Africa. J. Epidemiol. Community Health 54, 381–387. Pitt, M., Rosenzweig, M., 1990. Estimating the intrahousehold incidence of illness: child health and gender-inequality in the allocation of time. Int. Econ. Rev. 31 (4), 969–989. Ravallion, M., 2003. The debate on globalization, poverty and inequality: why measurement matters. Int. Aff. 79 (4), 739–753. Ravallion, M., Wodon, Q., 2000. Does child labour displace schooling? Evidence on behavioural responses to an enrollment. Econ. J. 110, 158–175. Rogers, S.G., 1983. Efforts toward women’s development in Tanzania: gender rhetoric vs. gender realities. In: Staudt, K., Jaquette, J. (Eds.), Women in Developing Countries: A Policy Focus. Haworth Press, New York, NY, pp. 23–41. Rose, E., 2000. Gender bias, credit constraints and time allocation in rural India. Econ. J. 110 (465), 738–758. Rosenzweig, M., Wolpin, K., 1993. Credit market constraints, consumption smoothing, and the accumulation of durable production assets in low-income countries: investment in bullocks in India. J. Polit. Econ. 101 (2), 223–244. Skoufias, E., 1993. Labor market opportunities and intrafamily time allocation in rural households in South Asia. J. Dev. Econ. 40, 277–310. Strauss, J., Thomas, D., 1995. Human resources: empirical modeling of household and family decisions. In: Behrman, J., Srinivasan, T. (Eds.), Handbook of Development Economics. Elsevier Science, Amsterdam, pp. 1884–2023. Sun, A., Yao, Y., 2010. Health shocks and children’s school attainments in rural China. Econ. Educ. Rev. 29, 375–382. Szekely, M., Hilgert, M., 1999. What’s Behind the Inequality We Measure: An Investigation Using Latin American Data. Inter-American Development Bank, Washington, D.C. Thomas, D., 1994. Like father, like son; like mother, like daughter: parental resources and child height. J. Human Resour. 29 (4), 950–988. Thomas, D., Beegle, K., Frankenberg, E., Sikoki, B., Strauss, J., Teruel, G., 2004. Education in a crisis. J. Dev. Econ. 74, 53–85. Thomas, D., Strauss, J., Henriques, M., 1991. How does mother’s education affect child height? J. Human Resour. 26 (2), 183–211. Townsend, R., 1994. Risk and insurance in village India. Econometrica 62 (3), 539–591. Wagstaff, A., 2007. The economic consequences of health shocks: evidence from Vietnam. J. Health Econ. 26, 82–100. Wagstaff, A., Lindelow, M., 2014. Are health shocks different? Evidence from a multishock survey in Laos. Health Econ. 23 (6), 706–718. World Bank, 1999. Tanzania Social Sector Review. The World Bank, Washington, D.C. World Bank, 2004. User’s guide to the Kagera health and development survey datasets. In: Working Paper. World Bank. World Bank, F.I., 2009. World Bank Gender in Agriculture Sourcebook. IBRD, Washington, D.C.